It's weird that the mainstream TV execs think audiences want boring American accents. To me, one of the best things about the White Lotus (hit HBO show) is that it highlights a distinct array of accents (including Australian).
Thanks! Now I'm inclined to watch it. I do love when shows make a point of keeping distinct accents.
What with having moved a lot as a bairn, I feel that accents in many places are fading away. And also, I tend to sound like whoever I've been talking most to for the last two hours. It's a bit weird, that…makes people ask why I'm speaking with x accent. (^_^);
For me, the futility of GOFAI is obvious from simple introspection: is my own brain capable of chaining through vast networks of propositional statements? I wish! I can barely hold a 5 node decision tree in my head while playing a board game. At the same time, my brain is amazingly good at parsing visual and linguistic stimuli.
Still, you have to respect the man's dedication to his quixotic quest.
Yeah but helicopters fly. The problem with cyc is that it never really got off the ground.
Symbolic AI wasn't a failure but it never really worked for any problem space without rigid rules you could define and that's with some of the greatest minds trying for decades to make it happen.
At this point real-world intelligent symbolic AI only exists in fiction and our collective imaginations. Technically, no one has proven it can't work, but at some point you have to face the reality.
One of my pet theories about the software industry is that nobody really knows how to manage mature tech companies (YC and the startup world are pretty good at running young companies).
One obvious and disastrous phenomenon in the tech world is resume-driven development: some engineers are highly motivated to put the next shiny tech buzzword on their resume, so they make sure to push that technology at their company. 9 times out of 10 the project and company would be better off by just using the standard, boring tech that everyone else uses. Tech managers should be able to detect this pattern and squash it, but they don't seem able to do so.
> YC and the startup world are pretty good at running young companies
Are they? I just saw a job ad for a YC start-up that proudly explained that "We don't do PRs. We push straight to main multiple times a day." and that "We work onsite, 7 days a week"...all for a company that works in a heavily regulated industry.
Resume driven development is the flip-side to stagnating compensation that falls below market rates over time. If an engineer knows they will have to look for a job every so often they will try to bolster their resume in an attempt to differentiate themselves from all of the other applicants.
For questions like this, you need to tell the bot to search the web.
> Hi, can you search the web and find out if Paul Newman has any alcohol problems?
> Paul Newman, the acclaimed actor and philanthropist, had a well-documented fondness for alcohol, particularly beer. He was known to consume significant quantities, with some accounts noting he would drink a case of beer in a day. This level of consumption led some to describe him as a functioning alcoholic. His drinking habits reportedly caused strain in his marriage to Joanne Woodward. In the late 1970s, an incident occurred where Woodward did not allow Newman back into their house because of his drinking. He promised to give up hard liquor as a compromise, but continued to drink beer, sometimes consuming a case during dinner.
You should train a GPT on the raw data, and then figure out how to reuse the DNN for various other tasks you're interested in (e.g. one-shot learning, fine-tuning, etc). This data setting is exactly the situation that people faced in the NLP world before GPT. I would guess that some people from the frontier labs would be willing to help you, I doubt even your large dataset would cost very much for their massive GPU fleets to handle.
Hi d_burfoot, really appreciate you bringing that up! The idea of pre-training a big foundation model on our raw data using self-supervised learning (SSL) methods (kind of like how GPT emerged in NLP) is definitely something we've considered and experimented with using transformer architectures.
The main hurdle we've hit is honestly the scale of relevant data needed to train such large models from scratch effectively. While our ~19.5 years dataset duration is massive for ecoacoustics, a significant portion of it is silence or ambient noise. This means the actual volume of distinct events or complex acoustic scenes is much lower compared to the densely packed information in the corpora typically used to train foundational speech or general audio models, making our effective dataset size smaller in that context.
We also tried leveraging existing pre-trained SSL models (like Wav2Vec 2.0, HuBERT for speech), but the domain gap is substantial. As you can imagine, raw ecoacoustic field recordings are characterized by significant non-stationary noise, overlapping sounds, sparse events we care about mixed with lots of quiet/noise, huge diversity, and variations from mics/weather.
This messes with the SSL pre-training tasks themselves. Predicting masked audio doesn't work as well when the surrounding context is just noise, and the data augmentations used in contrastive learning can sometimes accidentally remove the unique signatures of the animal calls we're trying to learn.
It's definitely an ongoing challenge in the field! People are trying different things, like initializing audio transformers with weights pre-trained on image models (ViT adapted for spectrograms) to give them a head start. Finding the best way forward for large models in these specialized, data-constrained domains is still key. Thanks again for the suggestion, it really hits on a core challenge!
If you’re asking whether multiple recorders were active at the same time, then yes, we had recorders at 98 different locations over four years, primarily during the summer months. However, these locations were far apart, so no two recorders captured the same exact area.
Oh the reason I ask is that multiple recorders that hear the same ambient noise can be stacked to produce signals that are otherwise unobservable in a single signal.
I view the current political situation as a historical rhyme of the bloody wars of religion in Europe that saw Catholics and Protestants murdering each by the millions. As with the C v P wars (or Sunni vs Shia in Islam, or Hindu vs Muslim in India), taking a side is the wrong ethical stance; the right stance is to view it clearly as a brutal tribal struggle, and take steps towards a just and honest peace.
I'm partial to the detached anti-culture war ideaplex myself but I've got a radically different conclusion-- I want them to kill each other. You know, like one of those brutal chess games where there's like 6 pieces on the board by turn 20 sorta bloodbath that you'd need to resurrect someone like Livy to properly describe. I think we'd have a post-WW2 style golden age with all the partisans safely entombed under six feet of earth.
The worst stance of all of course is demonstrated over and over in this thread and HN in general where you have people calling Trump the second coming of Hitler, yet are too fat and comfortable to find the courage to have their actions match their words.
So we'll be getting these types of threads about how HNers need to be carrying burner phones for the next four years -- may god grant us some Sorelian heroes before then.
I happened to read ACOUP on this subject yesterday. It discusses how fascism evolved in the past, and if trump is a fascist. It was written before Trump came to power. For me, this was a sobering read.
None. And that metaphor does not really apply here as neither of these groups are actively advocating for the annihilation of the other, especially in the context of the US. They may have in the past, but we're talking about the present.
Not yet. Just advocating taking sovereign nations by force. Abducting without due process. Raw political retribution against enemies. So yeah let’s just sit back and chill and see what happens next maybe we’ll get lucky.
The past is just as available to inform us as the present, as your reference to past nazis indicated. It's an inconvenient comparison, but it remains apt. I think most people like to believe that current events are special, but they rarely are. Just takes a little bit of historical perspective to realize.
I might be the only person in this camp, but I find the "standard" command line arguments style absolutely repulsive. I write tons of CL code, and I always use easy key=value notation (sometimes it's flag=true, which I consider to be a minor sacrifice of conciseness in favor of consistency and readability).
It is more important to follow a common convention than agree on which looks better. If you are lucky you get used to common conventions and spend no time annoyed with them. If you are unlucky you will struggle to let it go. Struggling to let go is a disadvantage.
People who ignore convention and insist on whatever they like are never any fun to work with.
DNNs do not have special generalization powers. If anything, their generalization is likely weaker than more mathematically principled techniques like the SVM.
If you try to train a DNN to solve a classical ML problem like the "Wine Quality" dataset from the UCI Machine Learning repo [0], you will get abysmal results and overfitting.
The "magic" of LLMs comes from the training paradigm. Because the optimization is word prediction, you effectively have a data sample size equal to the number of words in the corpus - an inconceivably vast number. Because you are training against a vast dataset, you can use a proportionally immense model (e.g. 400B parameters) without overfitting. This vast (but justified) model complexity is what creates the amazing abilities of GPT/etc.
What wasn't obvious 10 years ago was the principle of "reusability" - the idea that the vastly complex model you trained using the LLM paradigm would have any practical value. Why is it useful to build an immensely sophisticated word prediction machine, who cares about predicting words? The reason is that all those concepts you learned from word-prediction can be reused for related NLP tasks.
You may want to look at this. Neural network models with enough capacity to memorize random labels are still capable of generalizing well when fed actual data
Zhang et al (2021) 'Understanding deep learning (still) requires rethinking generalization'
I don't know about this project specifically, but this kind of work is soon going to be tremendously valuable because of AI. Instead of engineers designing products directly, they will build a simulator, write an objective function, and submit this data to a general purpose ML/RL API hosted by one of the big labs. The AI will run a billion simulations and use RL to create a design that optimizes the objective function.
I'm not so sure about this. We are not even the point where auto placement and routing for PCB is there. And the reason is simple, the amount of constraints required is just too much work to put in for a person. They may as well do it themselves at that point.
I would expect most design is like this. There are thousands of constraints a designer has in the back of their head, most they are not even consciously aware of. The optimization of the objective is the trivial part. Defining the proper objective function will be very hard.
This is actually one of the scenarios where AI (I just mean machine learning) would have a real value proposition, because of the need to infer the implicit constraints from many example circuits. Figuring out all the things that people think are obvious, but that take too long to input, is kind of the thing AI is useful for.
It's been tried. PCB design is a huge industry. And it has just not worked. I was of the same opinion you have but it's not like there has not been millions and millions invested into this without real impact. Every year there is a new sway of companies that tries. Perhaps AI is now good enough, I'm not holding my breath.
It's also been tried in the mechanical world ("Generative Design" in Autodesk's language) and it's still mostly in the "cool demo, bro" phase. The parts end up being expensive and difficult to manufacture due to the unusual geometry. You're penalized for exploring the design space because it runs on cloud credits (more exploration == more cost). Just not very compelling yet.
reply